A comparison and an exploration of the forecasting accuracy of a loglinear model at different levels of aggregation

We compare the scan*pro model of retail promotion effects, at different levels of aggregation. The alternative model specifications are: (1) store-level models with homogeneous or heterogeneous response parameters across retail chains, and with or without weekly indicator variables, (2) chain-level models with homogeneous or heterogeneous response parameters across retail chains, and (3) a market-level model. Based on scanner data, we show comparisons between the models in terms of relative frequencies of statistically significant parameter estimates in the expected range of values. Sales forecasts are compared at two levels viz. chain and market level. We find that a comparison of the relative frequencies favors the homogeneous store models (with or without weekly indicators), while the forecasting accuracy examined at both the chain and market levels is superior for chain-specific store models without weekly indicator variables. We also examine differences in the mean squared error between the estimation and validation samples.